Binary Simulated Normal Distribution Optimizer for feature selection: Theory and application in COVID-19 datasets.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Machine learning classification accuracy is highly dependent on the selected features.
• Many feature extraction methods yield irrelevant or redundant features, impacting performance.
• Feature selection (FS) is crucial for optimizing models by removing suboptimal features.

Purpose of the Study:

• To introduce a novel hybrid meta-heuristic algorithm, Binary Simulated Normal Distribution Optimizer (BSNDO), for effective feature selection.
• To enhance classification accuracy and model efficiency by optimizing feature subsets.
• To validate the proposed BSNDO method on diverse and high-dimensional real-world datasets.

Main Methods:

• A hybrid approach combining the Generalized Normal Distribution Optimizer (GNDO) with Simulated Annealing (SA) was developed.
• The proposed Binary Simulated Normal Distribution Optimizer (BSNDO) incorporates SA as a local search mechanism.
• The BSNDO method was evaluated on 18 UCI datasets, high-dimensional microarray datasets, and a COVID-19 classification dataset.

Main Results:

• The BSNDO method demonstrated superior performance compared to its predecessor and other popular feature selection techniques.
• Evaluations on UCI, microarray, and COVID-19 datasets confirmed the effectiveness of BSNDO.
• The proposed method achieved higher classification accuracy and improved efficiency.

Conclusions:

• The Binary Simulated Normal Distribution Optimizer (BSNDO) is a highly effective feature selection method.
• BSNDO offers significant improvements in classification accuracy and computational efficiency.
• The method shows promise for real-world applications, including medical data analysis.